Nutrition methodology

How close our estimates come

We test Capseta’s nutrition estimates against recipes whose publishers print their own nutrition analysis, and we publish the results with each nutrition release: how close we come, what we’re working to improve, and a dated record of what’s changed. How each figure is made is on our methodology page.

On this page

At a glance

How close it comes today

For each recipe we compare our calories with the publisher’s figure. The typical gap is the middle one: half the recipes come closer, half are further off. (Researchers call it the median.)

  • 11% typical calorie gap on 29 unseen test recipes, and it doesn’t lean high or low
  • 22 of 29 test recipes within 20% of the publisher’s calories, 13 of them within 10%
  • 13 other test recipes we couldn’t fully count yet. Counting those as misses, 22 of 42 come within 20%.

With 29 recipes, chance alone moves a share like this a lot: the true share within 20% could be anywhere from about 58% to 88% (a 95% interval). Coverage is not yet where it should be: we fully counted 29 of the 42 test recipes that don’t depend on a store brand (69%), and our bar is 80%.

On the 148 practice recipes, which we use to find and fix problems, the typical gap is 11%, and 111 of 148 come within 20%.

“Unseen” means we never read the test recipes one by one, the way a teacher keeps the final exam out of the practice questions. We drew them at random, half of each publisher’s recipes, before the method ever ran on them, and we open their results once per method version. They replaced an earlier test set on November 3, 2026, because we had used that set’s totals when choosing two fixes: how we count meat fat, and how we weigh plain salt. Those recipes are practice recipes now. The test recipes are the harder test and the one we lead with, and they show where we fall short: 3 of 29 are off by more than 35%. See what we’re working to improve.

Measured November 3, 2026, method version 0.21.1.

Back to top

The tests

How we test it

We compare our figures with recipes whose publishers print their own nutrition analysis for the same ingredient list. Some publishers say what software or database they used; most don’t. Their figures are calculations too, not lab tests, and they include the publisher’s own choices, such as fat drained off or the size of a serving. So this measures how closely we agree with published analyses. It doesn’t prove what’s in any dish.

We test ten nutrients: calories, protein, carbohydrate, fat, saturated fat, total sugars, fiber, sodium, potassium, and phosphorus (phosphorus on practice recipes only, because none of the test recipes’ publishers print it). The panel also shows trans fat, cholesterol, calcium, iron, vitamin D, and alcohol. We haven’t tested those against published figures yet, so treat them as rougher than the ones on this page.

We keep two groups of recipes:

  • Practice recipes. We read these recipes and use them to find and fix problems, so they’re the easier test, like practice questions you’ve already seen. (Researchers call this a development set.)
  • Test recipes. Kept unseen, like a final exam: we never read them one by one, only their totals. We chose them by a recorded random draw before the method ever ran on them. (Researchers call this a held-out set.)

Not every recipe we collect gets a score. We set a recipe aside when the publisher’s own figures don’t add up, when we can’t count every ingredient yet, or when its figures depend on a store-bought brand we can’t know. Scoring those would make our numbers look better or worse for the wrong reason.

We set the bar for each nutrient on October 28, 2026, before the first full run, and for saturated fat and total sugars on November 3, 2026, before either was first scored. We don’t loosen a bar to pass. A recipe counts as close when our figure is within the tolerance in the table, or within a small fixed amount for small figures, such as 25 calories, so rounding alone doesn’t fail it. A nutrient meets our bar when enough recipes are close and our figures don’t lean too far high or low. The tolerances are our own convention, not a regulatory standard: FDA’s 20% applies to lab tests of packaged foods against their labels, in one direction. We borrowed its size, not its rule.

Why each bar is where it is
Our bar for each nutrient
NutrientClose when withinOr withinShare that must be closeLean allowedWhy
Calories20%25 calories80%10%The headline figure. Careful analyses of the same ingredient list usually agree within this.
Protein20%2 g75%10%Driven by a few heavy ingredients, such as meat, beans, and dairy, whose weights recipes usually state.
Carbohydrate20%3 g75%10%Depends on dry versus cooked weights for grains and pasta, the most common weighing question.
Fat30%2 g70%15%Swings with cooking spray, oil, fat drained off, and how lean the meat is. Small figures make percentages jumpy.
Fiber30%1.5 g70%15%Fiber differs between lab methods and between databases, and publishers often round small amounts to whole grams.
Sodium35%75 mg65%15%Broth, canned goods, cheese, and sauces vary severalfold by brand, and salt “to taste” has no amount. Because sodium matters so much to many readers, the lean we allow stays at 15%.
Potassium30%100 mg70%15%Varies by variety and with what leaches into cooking water, which some publishers allow for and we don’t yet.
Phosphorus30%50 mg70%15%Phosphorus added in processing often isn’t in food tables, and few publishers report it.
Saturated fat30%1 g70%15%Follows fat: it swings with how lean the meat is, which dairy, and which oil, details recipes often leave out. Smaller figures than total fat, so a smaller fixed amount. Set November 3, 2026, before it was first scored.
Total sugars30%2 g70%15%Compared with the publisher’s total sugars, never added sugars, which we can’t calculate. Some USDA records have no sugar figure, so fewer recipes can be counted. Set November 3, 2026, before it was first scored.

We also require that at least 80% of the recipes we collect, apart from those set aside, can be fully counted, so the results can’t look better by leaving hard recipes out.

Practice recipes: 148 scored

  1. 196 collected
  2. −11 set aside: the publisher’s own figures don’t add up
  3. −23 not scored yet: we can’t count every ingredient
  4. −14 not scored: the figures depend on a store-bought brand
  5. 148 scored

Fully counted: 148 of the 171 that don’t depend on a store brand and weren’t set aside (87%). Our bar is 80%: meets it.

Share of recipes close to the published figure
  • Calories 75%, below our bar of 80%
  • Protein 76%, meets our bar
  • Carbohydrate 78%, meets our bar
  • Fat 78%, meets our bar
  • Fiber 85%, meets our bar
  • Sodium 77%, meets our bar
  • Potassium 71%, meets our bar
  • Phosphorus 82%, meets our bar
  • Saturated fat 76%, meets our bar
  • Total sugars 86%, meets our bar
Show the numbers
Practice recipes, measured November 3, 2026
NutrientClose to the published figure95% intervalClose on the percentage aloneTypical gapOur figures runOur bar
Calories 111 of 148 (75%) 67–81% 109 of 148 11% (23 calories) About 4% high Not yet
Protein 113 of 148 (76%) 69–82% 102 of 148 11% (1.1 g) Neither Meets it
Carbohydrate 113 of 145 (78%) 71–84% 103 of 145 11% (2.0 g) Neither Meets it
Fat 115 of 148 (78%) 70–84% 105 of 148 17% (1.2 g) Neither Meets it
Fiber 123 of 145 (85%) 78–90% 88 of 145 18% (0.5 g) Neither Meets it
Sodium 114 of 148 (77%) 70–83% 107 of 148 17% (43 mg) About 4% high Meets it
Potassium 67 of 94 (71%) 61–79% 66 of 94 16% (78 mg) About 11% high Meets it
Phosphorus 37 of 45 (82%) 69–91% 37 of 45 7% (24 mg) Neither Meets it
Saturated fat 97 of 128 (76%) 68–82% 70 of 128 26% (0.5 g) About 8% high Meets it
Total sugars 68 of 79 (86%) 77–92% 53 of 79 17% (0.8 g) Neither Meets it
How far off, recipe by recipe
NutrientWithin 10%10–20%20–35%More than 35%
Calories69402118
Protein71312323
Carbohydrate71322417
Fat55243534
Fiber41322537
Sodium55272541
Potassium35171923
Phosphorus26757
Saturated fat30232248
Total sugars3191518
See all 196 practice recipes

Some recipes count as close only because of the fixed amount for small figures: fiber, 35 of the 123; saturated fat, 27 of the 97; total sugars, 15 of the 68.

Test recipes: 29 scored

  1. 51 collected
  2. −13 not scored yet: we can’t count every ingredient
  3. −9 not scored: the figures depend on a store-bought brand
  4. 29 scored

Fully counted: 29 of the 42 that don’t depend on a store brand (69%). Our bar is 80%: not yet.

Share of recipes close to the published figure
  • Calories 76%, below our bar of 80%
  • Protein 72%, below our bar of 75%
  • Carbohydrate 72%, below our bar of 75%
  • Fat 83%, meets our bar
  • Fiber 79%, meets our bar
  • Sodium 62%, below our bar of 65%, and runs about 17% high, more than the 15% we allow
  • Potassium 85%, meets our bar
  • Saturated fat 86%, close often enough, but runs about 19% high, more than the 15% we allow
  • Total sugars 70%, meets our bar
Show the numbers
Test recipes, measured November 3, 2026
NutrientClose to the published figure95% intervalClose on the percentage aloneTypical gapOur figures runOur bar
Calories 22 of 29 (76%) 58–88% 21 of 29 11% (29 calories) Neither Not yet
Protein 21 of 29 (72%) 54–85% 19 of 29 15% (2.1 g) About 5% low Not yet
Carbohydrate 21 of 29 (72%) 54–85% 20 of 29 8% (2.1 g) Neither Not yet
Fat 24 of 29 (83%) 65–92% 21 of 29 14% (1.2 g) About 9% high Meets it
Fiber 23 of 29 (79%) 62–90% 20 of 29 17% (0.8 g) About 4% high Meets it
Sodium 18 of 29 (62%) 44–77% 16 of 29 31% (117 mg) About 17% high Not yet: runs about 17% high
Potassium 11 of 13 (85%) 58–96% 11 of 13 16% (91 mg) About 13% high Meets it
Saturated fat 24 of 28 (86%) 69–94% 18 of 28 22% (0.5 g) About 19% high Not yet: runs about 19% high
Total sugars 7 of 10 (70%) 40–89% 6 of 10 20% (0.8 g) About 10% high Meets it
How far off, recipe by recipe
NutrientWithin 10%10–20%20–35%More than 35%
Calories13853
Protein11846
Carbohydrate15545
Fat10676
Fiber10748
Sodium72712
Potassium3622
Saturated fat54116
Total sugars4123

Some recipes count as close only because of the fixed amount for small figures: fat, 3 of the 24; fiber, 3 of the 23; saturated fat, 6 of the 24.

Not shown: phosphorus, because none of these recipes’ publishers print that figure. We test it on the practice recipes.

We don’t list the test recipes, so they stay a fair test.

“Typical gap” is the median difference from the published figure, as a percentage and in the nutrient’s own units, per serving. “95% interval” is the range the true share could fall in, given how few recipes there are. “Close on the percentage alone” leaves out the fixed amount for small figures. “Our figures run” is the median direction: high, low, or neither. Potassium, phosphorus, saturated fat, and total sugars are scored only where the publisher printed them. Saturated fat and total sugars are also scored only where we had a figure for every ingredient.

Who published the recipes

Practice recipes

  • U.S. Department of Agriculture, MyPlate Kitchen: 29
  • U.S. Department of Veterans Affairs, Nutrition and Food Services: 23
  • National Heart, Lung, and Blood Institute (NHLBI), National Institutes of Health: 22
  • Celebrate Your Plate, Ohio Association of Foodbanks (created by The Ohio State University Extension, SNAP-Ed): 19
  • American Diabetes Association: 12
  • DaVita Inc.: 11
  • USDA Agricultural Research Service, Food and Nutrient Database for Dietary Studies (FNDDS) 2021–2023, via FoodData Central: 10
  • NHS (England), Healthier Families: 10
  • American Kidney Fund: 9
  • Academy of Nutrition and Dietetics: 9
  • Iowa State University Extension and Outreach: 9
  • Institute of Child Nutrition (ICN); USDA standardized recipe: 7
  • University of Illinois Extension: 6
  • Oregon State University Extension Service: 5
  • Heart Foundation of New Zealand: 4
  • Dietitians of Canada: 4
  • National Health and Medical Research Council (NHMRC), Australian Government: 3
  • National Diabetes Education Program (NDEP), National Institutes of Health and Centers for Disease Control and Prevention, US Department of Health and Human Services: 3
  • Leah's Pantry (EatFresh.org): 1

Test recipes

  • U.S. Department of Veterans Affairs, Nutrition and Food Services: 23
  • Celebrate Your Plate, Ohio Association of Foodbanks (created by The Ohio State University Extension, SNAP-Ed): 20
  • National Health and Medical Research Council (NHMRC), Australian Government: 4
  • National Diabetes Education Program (NDEP), National Institutes of Health and Centers for Disease Control and Prevention, US Department of Health and Human Services: 4

Naming a publisher doesn’t mean it endorses Capseta. The FNDDS recipes are USDA’s own mixed dishes, built from the same data we use, so they test our reading and weighing, not the nutrient values.

Back to top

Status

Where it stands

What we’re working to improve

Where our figures are weakest today, from the test results and from what the method leaves out. We update this list with each test run, and when something comes off it, What’s changed says when and what it did to the results.

As of November 3, 2026

In the test results

  • Calories, protein, carbohydrate, sodium, and saturated fat on the test recipes. They don’t meet our bar yet. We work on them by kind of error, using examples from the practice recipes.
  • Fat poured off, skimmed, or left on a grill. Recipes rarely say how much fat leaves the dish, so we count it as eaten. Bacon the recipe calls cooked is the one exception so far.
  • Potassium runs high. Many publishers allow for potassium lost into cooking water, and we don’t.
  • Recipes we can’t count fully. A food we can’t match yet keeps a recipe out of the scores. We fully counted 69% of the test recipes against a bar of 80%.

In the method

  • Little adjustment for cooking. We don’t adjust for water, vitamins, or fat lost in cooking, except alcohol, which counts only what survives. We do count the oil food soaks up in deep-frying, leave out what the recipe drains or takes out, count whole cuts of meat trimmed, and count bacon the recipe calls cooked as cooked bacon.
  • Phosphorus and potassium may read low for processed ingredients, because food makers often don’t report what they add.
  • USDA data describes American foods. Foods and brands from other countries may match less well, even where we use European tables for a few of them.
  • An “Estimated match” may be wrong. Check it if a figure surprises you, and tell us.

What it doesn’t do

  • Ten nutrients tested, not all of them. We’ve tested calories, protein, carbohydrate, fat, saturated fat, total sugars, fiber, sodium, potassium, and phosphorus (phosphorus on practice recipes only, because none of the test recipes’ publishers print it). The panel also shows trans fat, cholesterol, calcium, iron, vitamin D, and alcohol, which we haven’t tested against published figures yet, so treat them as rougher.
  • Allergens. We match foods to nutrition records, which can’t tell you what’s in the product you bought or what it was made alongside, so Capseta doesn’t check for allergens. If you or someone you cook for has an allergy, read the ingredient list on everything you use, especially packaged foods: brands differ, and manufacturers change their recipes. More on allergens
  • Total sugars only. The figure counts all the sugar in a dish, whether it comes from fruit and milk or from the sugar bowl. USDA’s records for whole foods report total sugars, not added sugars, so we can’t split them out.
  • No % Daily Value. Daily Values are set for a reference 2,000-calorie diet, not for you. Your needs depend on your age, size, activity, and health, so a percentage could mislead more than it helps. We show the amounts and leave the judgment to you and your care team.
  • Per serving means the recipe’s servings. We divide the whole recipe by the number of servings it gives. If you cut a lasagna into eight pieces when the recipe says six, each of your pieces holds less than the figure says. Pay attention to how you divide the finished dish, and change the servings in the panel to match.

What’s changed

When we change the method, the figures in your recipes can change too. Each change is listed here with its date and what it did to the test results, measured on the same recipes before and after.

  1. November 3, 2026

    Ingredient words reviewed. We reviewed the ingredient words the practice recipes brought in. Kitchen tools listed with the ingredients no longer count as food; store products whose figures depend on the brand, such as a branded seasoning, now ask you for the label; and dried chiles and chile powders share one set of figures. Every change came from practice recipes; the test recipes stayed unseen. On the 27 test recipes fully counted both times, calories close to the published figure stayed at 74%.

  2. November 3, 2026

    A new set of test recipes. We drew a fresh set from four publishers whose terms allow copying: the US Department of Veterans Affairs, Celebrate Your Plate, the National Diabetes Education Program, and Australia’s Eat For Health. The earlier test recipes had informed two fixes, so they became practice recipes. Its first results can’t be compared with earlier ones.

  3. November 3, 2026

    Saturated fat and total sugars tested. We now test both against the publishers’ figures, with bars we set before scoring them. The other figures didn’t change.

  4. November 2, 2026

    Meat as it’s eaten. We now count roasts, chops, and steaks trimmed of their outer fat unless the recipe says the fat stays, bacon the recipe calls cooked as cooked bacon, and ground chuck as 80% lean. Other nutrients stayed about the same.

  5. November 2, 2026

    First results published on this page.

Tell us

We publish this method while it’s still improving, because we want to hear where it falls short. Use “Tell us what looks wrong” in any recipe’s nutrition panel, or write to Hello@Capseta.app. A person reads every note, and when we had it wrong, the fix reaches everyone.

Back to top